GenAI in Telcommunications

Select papers from IEEE ICC 2025 Tutorial/Workshop

Paper 01 : NetOrchLLM

  • LLM as Intelligent router
  • LLM as Universal translator
  • NetOrchLLM Framework
    • LLM acts as coordinator, understanding intent and picking the right expert model.
    • Model repository holds the trusted analytical functions for wireless tasks.
    • Heavy computation is delegated to these specialized models, not the LLM.
  • In essence : Unstructured to structured repsentation translator

Ref : https://arxiv.org/abs/2412.10107

Paper 01 : NetOrchLLM

Paper 02 : ENWAR

flowchart LR
  A[Sensory Inputs] --> B[Textual Representation] --> C[Reasoning] --> D[Structured Output] --> E[V2X App] --> F[Wireless Action]

Ref : https://arxiv.org/abs/2410.18104

Paper 02 : ENWAR

Common Theme

Traditional Wireless System (Without LLM)

flowchart TD
    A["Raw Sensory Data<br/>(GPS, Camera, LiDAR, Radio)"] --> B["Preprocessing/Feature Extraction"]
    B --> C["Math/ML-based Algorithm<br/>(Beam Tracking, V2X Decisions, etc.)"]
    C --> D["Structured Output/Action<br/>(Beam Angle, Network Command)"]
    D --> E["Wireless System Actuation or V2X Application"]

LLM-Based Wireless System (ENWAR/NetOrchLLM Style)

flowchart TD
    A["Raw Sensory Data<br/>(GPS, Camera, LiDAR, Radio)"] --> B[Textualization<br/>Multi-Modal Embedding]
    B --> C["RAG/Knowledge Retrieval<br/>+ Context Fusion"]
    C --> D["LLM Reasoning<br/>(Interpret + Explain + Decide)"]
    D --> E["Convert Decision<br/>to Structured Action"]
    E --> F["Math/ML-based Algorithm<br/>(Precise Calculation)"]
    F --> G["Structured Output/Action<br/>(Beam Angle, Network Command)"]
    G --> H[Wireless System Actuation or V2X Application]

Paper 03 : NF Orch using LLMs

  • 3 Agents (Router, Resolver, Translator)
  • Each agent backed by KB store.
  • Actual deployment is still by traditional tools

Highlights

  • Highlevel control is fixed
  • Handle unstructed inputs

Returning theme

  • Universal Translator / Router
  • Unstructured -> Sturctured

Paper 04: Interface On Demand

  • 2 Agents (Match & CodeGen)
  • Match
    • Maps user intent to closest interface (API, Func call etc)
    • Instructs CodeGen to create wrappers
  • CodeGen
    • Generate working code

Highlights

  • Instant Vendor interoperablity
  • On-Demand system integration
  • Spec to Code - a reality